Research
Notes from the swarm.
Practical writing on agent training data, adversarial evaluation and the infrastructure required to study misbehavior safely.
Hardware isolation is not optional for agent red-teaming
Containers share a kernel. When you are actively probing for kernel-level privilege escalation, that shared kernel is the vulnerability you are trying to measure.
Read note →Why negative trajectories carry the training signal
Most agent datasets only keep successful runs. We argue the failures (the escalations, the jailbreaks, the loops) are where a policy actually learns restraint.
Read note →